Motion4Motion: Motion Transfer Across Subjects at Inference

Motion4Motion introduces a training-free framework for transferring motion between character videos at inference time. Instead of relying on predefined human skeletons and skeleton-conditional training, it models the character’s motion flow directly from video. This design targets motion transfer across diverse subjects and species while preserving their distinctive motion styles. The authors report extensive experiments and new applications showing improvements over baselines, although the supplied abstract does not provide numerical results, benchmark details, or implementation specifics. A project page is available from the authors.